Read replicas
Applications running in other regions incur cross-region latency to read the latest data from the leader. If a little staleness for reads is acceptable for the applications running in the other regions, then Read Replicas is the pattern to adopt.
A read replica cluster is a set of follower nodes connected to a primary cluster. These are purely observer nodes, which means that they don't take part in the Raft consensus and elections. As a result, read replicas can have a different replication factor (RF) than the primary cluster, and you can have an even number of replicas.
The replica serves a read when the application uses follower reads. Set yb_read_from_followers and mark the transaction read only. With follower reads off, the read goes to the tablet leader in the primary cluster, so placing the replica nearby does not lower read latency.
Suppose you have an RF 3 cluster set up in us-east-1 and us-east-2, with leader preference set to us-east-1. And suppose you want to run other applications in us-central and us-west. The read latencies would be similar to the following illustration.
Improve read latencies
To improve read latencies, set up separate Read Replica clusters in each of the regions where you want to run your application and where a little staleness is acceptable.
With follower reads enabled, the application reads from the closest replica, including a node in the read replica cluster, instead of going cross-region to the tablet leaders.
Notice that the read latency for the application in us-west has drastically dropped to 2 ms from the initial 60 ms and the read latency of the application in us-central has also dropped to 2 ms.
As replicas may not be up-to-date with all updates, by design, the result is slightly stale. The default staleness is 30 seconds. Set the interval with yb_follower_read_staleness_ms.
Failover
When the read replicas in a region fail, the application will redirect its read to the next closest read replica or leader.
Notice how the application in us-west reads from the follower in us-central when the read replicas in us-west fail. In this case, the read latency is 40 ms, still much less than the original 60 ms.